Thx Junpeng!
Yes, they are known, but they vary from one observation to the other, so with your suggestion I guess I would need several p vectors, one for each configuration?
To be more precise, here is a sample of the data:
Categories 1, 3 and 4 are the more likely to be not available for choice (I don’t have all of the data yet but I know that). Category 7 is an aggregation of “all other choices”, so it’s always available - i.e the zeros are real multinomial zeros, while zeros in cat1…6 reflect non-availability.
As you can see, the non-available categories vary, so I don’t really know how to say that to the model… My goal is to model these data as a hierarchical regression, with subject_type as cluster (and probably date for a second-stage refinement).
Thank you very much in advance!